Actually measured gravity and magnetic data denoising method based on priori knowledge deep learning

Abstract

The invention provides a priori knowledge deep learning-based gravity and magnetic data denoising method, which comprises the following steps of: constructing a data set fused with priori knowledge, carrying out artificial filtering processing on gravity and magnetic data to obtain a denoising result, slicing the result to form a label, simulating to generate noise data, and superposing the noise data on the label, forming deep learning network model input data in one-to-one correspondence with the labels; building a deep learning network model; constructing a loss function fused with prior knowledge, and optimizing a prediction result of deep learning; carrying out training and hyper-parameter tuning on the deep learning network model; and inputting the gravity and magnetic data of the prediction area into the trained deep learning network model to obtain a deep learning prediction result, and carrying out optimization processing on the deep learning prediction result to obtain a final gravity and magnetic data denoising result. The method can effectively process noisy gravity and magnetic data obtained through aviation, shipborne and satellite platforms, and removes a certain amount of data generated by platform movement, instrument observation or surrounding environment interference and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of gravity and magnetic exploration, and particularly relates to a gravity and magnetic data denoising method based on deep learning with prior knowledge. Background Art

[0002] In the technical field of gravity and magnetic exploration, gravity and magnetic data obtained through aerial, shipborne, and satellite platforms all generate a certain amount of noise signals due to platform movement, instrument observation, or surrounding environmental interference, etc., which even mask useful information, seriously affecting the subsequent processing and interpretation applications of gravity and magnetic data. Artificial filtering methods such as the Gaussian filtering method for the selection of filtering methods and filtering parameters have certain empiricism, and the processing process is also relatively complex. Existing deep learning methods lack the introduction of prior knowledge. In the process of constructing the data set, most of them use forward simulation of underground geological bodies with simple structures to obtain labeled data, making the simulated labeled data relatively simple and unable to truly simulate gravity and magnetic data. Summary of the Invention

[0003] In view of the technical problems existing in the above background art, the present invention proposes a gravity and magnetic data denoising method based on deep learning with prior knowledge. Its concept is reasonable and can effectively process noisy gravity and magnetic data obtained through aerial, shipborne, and satellite platforms, removing a certain amount of noise signals generated due to platform movement, instrument observation, or surrounding environmental interference, etc., which even mask useful information, thereby providing high-quality basic data for the subsequent processing and interpretation applications of gravity and magnetic data.

[0004] To solve the above technical problems, a gravity and magnetic data denoising method based on deep learning with prior knowledge provided by the present invention mainly includes the following steps:

[0005] (1) Training stage

[0006] (1.1) Construct a data set integrating prior knowledge. On the basis of analyzing the noise characteristics and composition of gravity and magnetic data, perform artificial filtering on the gravity and magnetic data to obtain a denoising result, slice this result to form a label, and simulate and generate noise data according to the noise characteristics of the gravity and magnetic data, and superimpose it on the label to form the input data of the deep learning network model corresponding to the label one by one;

[0007] (1.2) Build a deep learning network model according to the gravity and magnetic data denoising requirements;

[0008] (1.3) Construct a loss function integrating prior knowledge. The harmonic field is a characteristic of the gravity field and has a smooth property. Add a Laplace constraint to the loss function to optimize the prediction result of deep learning;

[0009] (1.4) Train the deep learning network model and optimize the hyperparameters. Stop training after meeting the set number of iterations or when the loss function reaches the threshold, and save the deep learning network model;

[0010] (2) Prediction stage

[0011] (2.1) Input the gravity and magnetic data of the prediction area into the trained deep learning network model to obtain the prediction results of deep learning;

[0012] (2.2) Conduct fine optimization on the prediction results of deep learning to obtain the final denoising results of the gravity and magnetic data.

[0013] In the gravity and magnetic data denoising method based on deep learning with prior knowledge, in step (1.1), the specific process of constructing a dataset integrating prior knowledge is as follows: starting from the original noisy satellite gravity data, obtain a smooth denoising result by performing omnidirectional filtering on the noisy gravity data; then crop and generate sample label data according to the set window size and moving step; and then add simulated noise data to the corresponding sample labels according to the noise characteristics of gravity, thus forming a large number of sample sets containing sample data and label data.

[0014] In the gravity and magnetic data denoising method based on deep learning with prior knowledge, in step (1.1), select the original noisy satellite gravity data in the area with longitude 45° - 90° and latitude - 45° - 0°, with a grid spacing of 0.075°, perform omnidirectional filtering on it to obtain a smooth denoising result; use a moving window to crop the denoised grid, select the window size as the sample size of 160×160, and the moving step as 4 times the grid spacing to obtain 12,100 sample label data; for each sample label data, add the noise data simulated by the following formula (7):

[0015]

[0016] In the above formula (7), A is the amplitude of the simulated strip noise, κ is the wavelength, α is the direction angle, a is the wavelength difference, R 0 is Gaussian white noise, x is the abscissa of the noise data calculation point, and y is the ordinate of the noise data calculation point.

[0017] In the gravity and magnetic data denoising method based on deep learning with prior knowledge, in step (1.2), first construct the mapping relationship between the noisy data and the noise - free data by learning a large number of training data for denoising the original satellite gravity data. This mapping relationship can be expressed as:

[0018] d 2 = Net(d 1 ,θ) (1);

[0019] In the above formula (1), d 1 and d 2 are the input data and output data of the deep learning network respectively, and θ is the deep learning network parameter.

[0020] In the gravity and magnetic data denoising method based on prior knowledge deep learning, wherein: the deep learning network model has a total of 6 layers of structure, consisting of 25 two-dimensional convolutional layers, 5 downsampling layers, 5 upsampling layers and 5 skip connections; the input data d 1 and the output data d 2 both have a resolution of 160×160. The resolutions of the data in the 2nd to 6th layers are 80×80, 40×40, 20×20, 10×10, 5×5 respectively. On the contrary, the number of channels increases from 64 in the first layer to 2048 in the 6th layer in a 2-fold manner; and both the two-dimensional convolutional layer and the downsampling layer are composed of 1 two-dimensional convolutional operator, 1 batch normalization operator and 1 Leaky-ReLU activation function, and the upsampling layer is implemented through two-dimensional transposed convolution.

[0021] In the gravity and magnetic data denoising method based on prior knowledge deep learning, wherein the specific process of training the deep learning network model in step (1.4) is: during network training, the input data enters the network from Figure 1 the upper left corner of the network, and extracts the feature information of this layer scale through 2 consecutive convolutional operations in the first layer of the deep learning network; then performs downsampling operation to halve the resolution of the data and retain the feature information obtained in the upper layer; repeat the above operations until the last layer. During this process, the receptive field is continuously increased, and the feature extraction from local information to global information of the input data is gradually realized; then enter the upsampling operation on the right side, restore the resolution of the data layer by layer, and realize the information fusion with the same-resolution features on the left side through skip connections; finally, 5 two-dimensional convolutional layers are used at the output end of the network to obtain the denoising result of the noisy data.

[0022] In the gravity and magnetic data denoising method based on prior knowledge deep learning, wherein the specific process of step (1.3) is:

[0023] First, the Dice function is introduced into the gravity data denoising to represent the similarity or overlap degree between the denoised data and the original data. When the Dice function is 1, it means that the denoised data and the original data completely overlap. Therefore, the Dice loss function is constructed as:

[0024]

[0025] In the above formula (2), d are the predicted value of gravity data denoising and the theoretical value of gravity data denoising respectively, N is the number of training samples, and M is the number of grid points of each sample.i Let \(i\) be a variable indicating the change in the number of training samples, and \(j\) be a variable indicating the change in the number of grid points for each sample.

[0026] Furthermore, the Laplace operator is introduced to constrain the smoothness of the prediction results of the deep learning network model; the Laplace operator can be expressed as the predicted value of gravity data denoising On a two-dimensional grid, it is the sum of the two-dimensional partial derivatives in the \(x\) - direction and \(y\) - direction respectively:

[0027]

[0028] Then the Laplace loss can be expressed in the following discrete form:

[0029]

[0030] Therefore, the total loss function, which is the sum of the above - mentioned formula (2) and formula (4), can be expressed as:

[0031]

[0032] In the above formula (5), \(\lambda\) is a weight coefficient used to adjust the proportion between the Dice loss and the Laplace loss.

[0033] In the method for denoising gravity and magnetic data based on deep learning with prior knowledge, in step (2.2), the way to perform fine - tuning on the prediction results of deep learning is: by introducing Gaussian filtering to perform low - pass filtering on the prediction results of deep learning, so as to obtain the final gravity data denoising result, where the Gaussian kernel used is:

[0034]

[0035] Adopting the above technical solution, the present invention has the following beneficial effects:

[0036] The method for denoising gravity and magnetic data based on deep learning with prior knowledge in the present invention is reasonably conceived and can be widely applied in the field of gravity and magnetic exploration technology. It can effectively process the noisy gravity and magnetic data obtained through aerial, ship - borne, and satellite platforms, remove the noise signals generated due to platform movement, instrument observation, or surrounding environmental interference, etc., which may even mask useful information, so as to provide high - quality basic data for the subsequent processing and interpretation applications of gravity and magnetic data.

[0037] Compared with the artificial filtering method, the present invention does not need to select complex filtering parameters, and can avoid the artificial errors caused by the artificial selection of filtering parameters in the artificial filtering method; compared with the existing deep learning methods, the advantages are that prior knowledge is introduced from two aspects. One is to incorporate prior knowledge into the training samples, using the results of artificial filtering processing as labels, simulating and generating noise data according to the noise characteristics of gravity and magnetic data, and superimposing it on the labels to form the input data of the deep learning network model corresponding to the labels one by one; the other is to construct a loss function that integrates prior knowledge. According to the fact that the gravity field is a harmonic field and has smooth characteristics, Laplace constraints are added to the loss function to optimize the prediction results of deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0039] Figure 1 is the flowchart of the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention;

[0040] Figure 2 is the schematic diagram of the sample set label data clipping involved in the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention;

[0041] Figure 3 is the schematic diagram of the samples and labels involved in the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention ((a, b), (c, d), (a, b), (e, f) are the samples and labels with serial numbers 11, 200, 2100 respectively);

[0042] Figure 4 is the schematic diagram of the deep learning network model for gravity data denoising involved in the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention;

[0043] Figure 5 is the loss curve diagram of the training set and validation set involved in the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention;

[0044] Figure 6 is the schematic diagram of the validation data involved in the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention ((a) validation data (b) theoretical value of the denoised validation data (c) noise of the validation data);

[0045] Figure 7This is the verification data denoising result diagram involved in the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention ((a) U-net denoising result (b) Noise removed by U-net (c) U-net + GF denoising result (d) Noise removed by U-net + GF);

[0046] Figure 8 This is the Bouguer gravity data diagram of the test area on the moon involved in the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention;

[0047] Figure 9 This is the comparison diagram of the denoising effect of real data involved in the gravity and magnetic data denoising method based on prior knowledge deep learning of the present invention ((a) ODF denoising result (b) Noise removed by ODF (c) Denoising result of the method of the present invention (d) Noise removed by the method of the present invention). Detailed implementation manners

[0048] Next, the technical solutions of the present invention will be described clearly and completely with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0049] Next, the present invention will be further explained and described in combination with specific implementation manners.

[0050] As Figure 1 shown, a gravity and magnetic data denoising method based on prior knowledge deep learning provided in this embodiment has the following specific process:

[0051] S100. Training stage

[0052] S101. First, construct a dataset integrating prior knowledge. On the basis of analyzing the noise characteristics and composition of gravity and magnetic data, use filtering methods such as Gaussian filtering to perform artificial filtering on gravity and magnetic data to obtain the denoising result based on the artificial filtering method. Slice this result at a certain step size to form a large number of labels. Simulate and generate noise data according to the noise characteristics of gravity and magnetic data, and superimpose it on the labels to form the input data of the deep learning network model corresponding to the labels one by one;

[0053] S102. Build a deep learning network model according to the gravity and magnetic data denoising requirements;

[0054] S103. Construct a loss function integrating prior knowledge. The harmonic field is a characteristic of the gravity field and has smooth characteristics. Add Laplace constraints to the loss function;

[0055] S104. Train the deep learning network model and tune the hyperparameters. Stop training after meeting the set number of iterations or when the loss function reaches the threshold, and save the deep learning network model.

[0056] S200. Prediction stage

[0057] First, input the gravity and magnetic data of the prediction area into the trained deep learning network model to obtain the prediction results of deep learning. Subsequently, perform subtle optimization on the deep learning prediction results through filtering methods such as Gaussian filtering to obtain the final denoising results of the gravity and magnetic data.

[0058] The present invention adopts the supervised deep learning method, and the data processing workflow is as Figure 1 shown:

[0059] In the training dataset stage, first, construct noise-free gravity and magnetic data as output labels. These labels are the results after denoising by artificial filtering methods, that is, the labels contain prior knowledge of the useful signals of the gravity and magnetic fields. Subsequently, the sum of the labels and the simulated gravity and magnetic noise is used as the input data of the deep neural network, that is, the input data contains the artificial empirical understanding of the noise. On this basis, fine-tune the hyperparameters to train the deep learning network model to obtain the optimal neural network parameters. In the prediction stage, use the trained network to predict the noisy gravity and magnetic data to obtain the denoised gravity and magnetic data. According to the smooth property of the gravity and magnetic fields as harmonic fields, perform optimization on the deep learning prediction results through artificial filtering methods to obtain the final results. The key to the above process is to develop a sufficient number of and highly accurate noisy gravity and magnetic datasets to comprehensively and objectively reflect the complexity of the gravity and magnetic noise and the characteristics of the gravity and magnetic fields.

[0060] Taking the denoising of noisy lunar gravity data as an example in this embodiment, the specific process of constructing the dataset integrating prior knowledge in the above step S101 is as follows:

[0061] Deep learning is a technology that simulates the learning process of the human brain by constructing and training neural network models. In this process, the construction of the dataset plays a crucial role. The construction of the deep learning dataset in this invention is different from that of previous scholars who started from the forward modeling of the theoretical model of underground geological bodies. Instead, it starts from the original noisy satellite gravity data. By performing omnidirectional filtering on the noisy gravity data in a certain area, a smooth denoising result is obtained. Then, a large number of sample label data are generated by cropping according to a certain window size and moving step. According to the noise characteristics of lunar gravity, simulated noise data are added to the corresponding sample labels, thus forming a large number of sample sets containing sample data and label data. The advantage of the above dataset construction is that the deep learning network model can not only learn artificial processing experience but also learn richer noise characteristics without regional restrictions. Therefore, in the prediction stage for other regions, a denoising result close to or even better than that of artificial denoising can be obtained.

[0062] Specifically, the original noisy satellite gravity data in the area with longitude 45° - 90° and latitude - 45° - 0° are selected in this invention, with a grid spacing of 0.075°. Omnidirectional filtering is performed on it to obtain a smooth denoising result (as Figure 2 shown). A moving window is used to crop the denoised grid. The window size is selected as the sample size (160×160), and the moving step is 4 times the grid spacing, obtaining 12,100 sample label data. For each sample label data, the following simulated noise data are added:

[0063]

[0064] where A is the amplitude of the simulated strip noise, κ is the wavelength, α is the direction angle, a is the wavelength difference, R 0 is Gaussian white noise, x is the abscissa of the noise data calculation point, and y is the ordinate of the noise data calculation point.

[0065] In this invention, A = 1, κ is randomly selected from integers between 3 and 6, a is randomly selected from integers between 1 and 8, α = 0 0 、30 0 、60 0 、90 0 、120 0 、150 0 represent 6-direction strip noises, and R 0 is Gaussian white noise with a standard deviation of 1 and a mean of 0. The sample label data are respectively summed with the simulated noise data to obtain 12,100 samples and corresponding labels. The first 10,100 are selected as the training set, and the last 2,000 are selected as the validation set. Figure 3Samples and labels with serial numbers 11, 200, and 2100 are given. The noisy sample data is on the left, and the noiseless label data is on the right.

[0066] The above-mentioned step S102 solves the denoising problem of the original satellite gravity data through deep learning technology. The core lies in constructing a complex mapping relationship between the noisy data and the noiseless data by learning a large amount of training data. This mapping relationship can be expressed as:

[0067] d 2 =Net(d 1 ,θ) (1);

[0068] where d 1 , d 2 are the input data and output data of the deep learning network model respectively, and θ is the parameter of the deep learning network model.

[0069] As Figure 4 shown, the present invention designs a deep learning network model based on the U-Net architecture for denoising the original satellite gravity data. The deep learning network model designed by the present invention has a total of 6-layer structure, consisting of 25 two-dimensional convolutional layers (kernel size 3×3, stride 1), 5 downsampling layers (kernel size 2×2, stride 2), 5 upsampling layers (kernel size 2×2, stride 2), and 5 skip connections. The red numbers represent the grid size of the data, and the black numbers represent the number of channels. The resolutions of the input data and output data are both 160×160, and the resolutions of the data in the 2nd to 6th layers are 80×80, 40×40, 20×20, 10×10, 5×5 respectively. On the contrary, the number of channels increases from 64 in the 1st layer to 2048 in the 6th layer in a 2-fold manner. Among them, the two-dimensional convolutional layer and the downsampling layer are both composed of 1 two-dimensional convolutional operator, 1 batch normalization operator, and 1 Leaky-ReLU activation function. The upsampling layer is implemented through two-dimensional transposed convolution.

[0070] The specific process of training the deep learning network model in the above-mentioned step S104 is as follows:

[0071] During network training, the input data enters the network from the Figure 1 upper left corner of, and extracts the feature information of this layer scale through 2 consecutive convolutional operations in the 1st layer. Then, a downsampling operation is performed to halve the resolution of the data and retain the feature information obtained in the upper layer. The above operations are repeated until the last layer. During this process, the receptive field is continuously increased, and the feature extraction from the local information to the global information of the input data is gradually realized. Subsequently, it enters the upsampling operation on the right side, restores the resolution of the data layer by layer, and realizes the information fusion with the same-resolution features on the left side through skip connections. Finally, 5 two-dimensional convolutional layers are used at the output end of the network to obtain the denoising result of the noisy data.

[0072] The specific process of the above step S103 is as follows:

[0073] In deep learning, the loss function plays a crucial role and is a non - negative real - valued function that measures the difference between the predicted output of the model and the true target value. The Dice function focuses on the overlapping part between the prediction result and the true label, which can better preserve the details and structure of the image. In this invention, it is introduced into gravity data denoising to represent the similarity or overlapping degree between the denoised data and the original data. When the Dice function is 1, it means that the denoised data and the original data completely overlap. Therefore, the Dice loss function is constructed as:

[0074]

[0075] Where d represents the predicted value of gravity data denoising and the theoretical value of gravity data denoising respectively, N is the number of training samples, M is the number of grid points for each sample, i is a variable indicating the change in the number of training samples, and j is a variable indicating the change in the number of grid points for each sample.

[0076] The Dice loss function can constrain the similarity degree between the denoised data and the original data, but it is difficult to achieve smooth transition between adjacent data. However, the Bouguer gravity anomaly caused by underground geological bodies has a smooth characteristic on the plane. Therefore, the Laplace operator is introduced to constrain the smoothness of the prediction result. The Laplace operator can be expressed as the sum of the second - order partial derivatives of the predicted value of gravity data denoising in the x - direction and y - direction on a two - dimensional grid:

[0077]

[0078] Then the Laplace loss can be expressed in the following discrete form:

[0079]

[0080] Therefore, the total loss function can be expressed as:

[0081]

[0082] Where λ is the weight coefficient, which is used to adjust the proportion between the Dice loss and the Laplace loss.

[0083] The verification of the data denoising result shows that there are still a small amount of high - frequency noises in the predicted value of gravity data denoising constrained by the Dice loss and the Laplace loss Therefore, Gaussian filtering is introduced to perform low - pass filtering on the deep - learning prediction result, so as to obtain the final gravity data denoising result. The Gaussian kernel used is:

[0084]

[0085] Result comparison and analysis:

[0086] After building the deep neural network architecture and preparing the training set and validation set, the training of the neural network can begin. After multiple tests and adjustments of the key hyperparameters, the learning rate is finally selected as 0.01, the batch size is 60, the dropout rate is 0.2, and the loss function threshold is 0.0001. The Adam algorithm is used in the present invention to optimize the parameters of the deep neural network. The computer configuration is AMD Ryzen 9 7945HX with Radeon Graphics 2.50GHz, RAM 16.0GB, and the PyTorch version is 2.1.2. The loss curves of the training set and validation set are as Figure 5 shown. In the initial stage of training, the losses of both the training set and the validation set are relatively large. However, as the number of iterations increases, the loss curves rapidly decay. Finally, the two loss curves tend to be stable and approach zero, without overfitting or underfitting phenomena.

[0087] Model data denoising results:

[0088] Figure 6 For the model data example, where a is the model data, b is the theoretical value (label) of the model data denoising, and c is the noise contained in the model data (the difference between a and b). Through the U-Net neural network learned by the present invention, Figure 7 the denoising result of a can be obtained, Figure 7 b is the noise removed by the U-Net neural network. Comparing Figure 6 b and Figure 6 c, most of the noise has been removed, but there are still a small part of high-frequency interferences. The result after Gaussian filtering optimization is shown in Figure 7 c. It can be seen by comparison that a satisfactory denoising result can be obtained by the denoising method combining deep learning and artificial filtering proposed by the present invention.

[0089] Real data denoising results:

[0090] Based on the satisfactory denoising results of the theoretical data, the original noisy satellite gravity data in the area of longitude 6° - 18° and latitude 0° - 12° ( Figure 8 ) is selected for testing, and the grid spacing is 0.075°. The comparison of the denoising results between the denoising method proposed by the present invention and the ODF method is shown in Figure 9 shown, where Figure 99a is the denoising result of the ODF method, 9b is the noise removed by the ODF method, 9c is the denoising result of the method of the present invention, 9d is the noise removed by the method of the present invention, and 9e is the curve comparison of the two methods on the AB section. From the plane and section results, both methods effectively remove the interference of strip noise, but the method of the present invention better retains some local anomaly information, such as Figure 8 shown by the red square in

[0091] The concept of the present invention is reasonable and can effectively process the noisy gravity and magnetic data obtained through airborne, shipborne and satellite platforms, removing the noise signals generated due to platform movement, instrument observation or surrounding environment interference, etc., which may even mask useful information, so as to provide high-quality basic data for the subsequent processing and interpretation applications of gravity and magnetic data.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A gravity and magnetic data denoising method based on deep learning of prior knowledge, characterized in that: The main steps include: (1) Training phase (1.1) Construct a data set that integrates prior knowledge. Based on the analysis of the noise characteristics and composition of gravity and magnetic data, perform artificial filtering on the gravity and magnetic data to obtain denoising results. Slice the results to form labels. Simulate and generate noise data based on the noise characteristics of gravity and magnetic data, and superimpose them on the labels to form input data for the deep learning network model that corresponds one-to-one with the labels. (1.2) Build a deep learning network model based on the denoising requirements of gravity and magnetic data; (1.3) Construct a loss function that integrates prior knowledge. The harmonic field is a characteristic of the gravity field and has smooth characteristics. Add Laplace constraints to the loss function to optimize the prediction results of deep learning. (1.4) Train the deep learning network model and tune the hyperparameters. Stop training when the set number of iterations is met or the loss function reaches a threshold, and save the deep learning network model. (2) Prediction stage (2.1) Input the gravity and magnetic data of the prediction area into the trained deep learning network model to obtain the prediction results of deep learning; (2.2) The prediction results of deep learning are slightly optimized to obtain the final denoising results of gravity and magnetic data.

2. The method for denoising gravity and magnetic data based on deep learning of prior knowledge as claimed in claim 1 is characterized in that ,The specific process of constructing a data set integrating prior knowledge in step (1.1) is as follows: starting from the original noisy satellite gravity data, a smooth denoising result is obtained by omnidirectional filtering of the noisy gravity data; then, the sample label data is generated by cropping according to the set window size and moving step size; and then, according to the noise characteristics of gravity, simulated noise data is added to the corresponding sample label, thereby forming a large number of sample sets containing sample data and label data.

3. The method for denoising gravity and magnetic data based on deep learning of prior knowledge as claimed in claim 2 is characterized in that: The step (1.1) is to select the original noisy satellite gravity data in the area of ​​longitude 45° to 90° and latitude -45° to 0°, with a grid spacing of 0.075°, and perform omnidirectional filtering on it to obtain a smooth denoising result; use a moving window to crop the denoising grid, select the window size as the sample size 160×160, and the moving step length as 4 times the grid spacing to obtain 12100 sample label data; for each sample label data, add the noise data simulated by the following formula (7): In the above formula (7), A is the simulated strip noise amplitude, κ is the wavelength, α is the direction angle, a is the wavelength difference, R0 is Gaussian white noise, x is the horizontal coordinate of the noise data calculation point, and y is the vertical coordinate of the noise data calculation point.

4. The method for denoising gravity and magnetic data based on deep learning of prior knowledge as claimed in claim 1, characterized in that: The step (1.2) is to first construct a mapping relationship between noisy data and noise-free data by learning a large amount of training data, which is used for denoising the original satellite gravity data. This mapping relationship can be expressed as: d2= Net(d1,θ) (1); In the above formula (1), d1 and d2 are the input data and output data of the deep learning network, respectively, and θ is the deep learning network parameter.

5. The method for denoising gravity and magnetic data based on deep learning of prior knowledge as claimed in claim 4, characterized in that: The deep learning network model has a total of 6 layers, consisting of 25 two-dimensional convolutional layers, 5 downsampling layers, 5 upsampling layers and 5 jump connections; the resolutions of the input data d1 and output data d2 of the deep learning network model are both 160×160, and the resolutions of the 2nd to 6th layer data are 80×80, 40×40, 20×20, 10×10, and 5×5, respectively. On the contrary, the number of channels is doubled from 64 in the 1st layer to 2048 in the 6th layer; and the two-dimensional convolutional layer and the downsampling layer are both composed of 1 two-dimensional convolution operator, 1 batch normalization operator and 1 Leaky-ReLU activation function, and the upsampling layer is implemented by two-dimensional transposed convolution.

6. The method for denoising gravity and magnetic data based on deep learning of prior knowledge as claimed in claim 1, characterized in that: The specific process of training the deep learning network model in the step (1.4) is as follows: during network training, the input data enters the network from the upper left corner of Figure 1, and extracts the feature information of the scale of this layer by entering the first layer of the deep learning network and performing two consecutive convolution operations; then a downsampling operation is performed to halve the resolution of the data and retain the feature information obtained in the upper layer; the above operation is repeated until the last layer, and the receptive field is continuously increased in this process, gradually realizing the feature extraction from the local information of the input data to the global information; then the upsampling operation on the right side is performed to restore the resolution of the data layer by layer, and realize the information fusion with the same resolution features on the left side through jump connections; finally, 5 two-dimensional convolution layers are used at the output end of the network to obtain the denoising result of the noisy data.

7. The method for denoising gravity and magnetic data based on deep learning of prior knowledge as claimed in claim 1, characterized in that: The specific process of step (1.3) is as follows: First, the Dice function is introduced into gravity data denoising to characterize the similarity or overlap between the denoised data and the original data. When the Dice function is 1, it means that the denoised data and the original data completely overlap, so the Dice loss function is constructed as: In the above formula (2), d is the predicted value of gravity data denoising, d is the theoretical value of gravity data denoising, N is the number of training samples, M is the number of grid points for each sample, i is a variable indicating the change in the number of training samples, and j is a variable indicating the change in the number of grid points per sample. The Laplace operator is then introduced to constrain the smoothness of the prediction results of the deep learning network model; the Laplace operator can be expressed as the denoised prediction value of the gravity data The sum of the two-dimensional partial derivatives in the x-direction and y-direction on a two-dimensional grid: The Laplace loss can be expressed in the following discrete form: Therefore, the total loss function, which is the sum of equation (2) and equation (4), can be expressed as: In the above formula (5), λ is a weight coefficient, which is used to adjust the ratio between Dice loss and Laplace loss.

8. The method for denoising gravity and magnetic data based on deep learning of prior knowledge as claimed in claim 1, characterized in that: The step (2.2) performs a subtle optimization process on the prediction results of deep learning by introducing a Gaussian filter to perform a low-pass filter on the prediction results of deep learning, thereby obtaining the final gravity data denoising result, wherein the Gaussian kernel used is:

Images